As of November 6, Reuters’ November 5 report revealed a previously unreported stockpile of about 600,000 Nvidia H200 chips in Hong Kong awaiting shipment to mainland China—not because buyers lacked demand, but because Beijing was withholding large-scale import approval.2 The report illustrates a two-sided squeeze: US export controls restrict which advanced chips Chinese firms can buy, while China’s own import restrictions limit access even to chips Washington has authorized.2

• US permission did not guarantee Chinese entry. Washington had authorized H200 sales to approved Chinese buyers following US–China trade talks in 2025, but Beijing remained reluctant to deepen dependence on US technology despite rising domestic AI demand.2

• Deliveries were limited, not entirely blocked. Sources said several major Chinese internet companies had each received about 10,000 H200s, with smaller quantities arriving through intermediaries and indirect routes—far short of the Hong Kong stockpile.12 Chinese regulators and Nvidia, Alibaba, ByteDance, DeepSeek and Tencent did not respond to Reuters’ requests for comment, leaving the account without their public confirmation.2

• Unofficial-market prices showed the scarcity premium. An eight-card H200 server was offered for more than 6 million yuan, approximately $900,000 and roughly triple its price a year earlier; export-controlled Blackwell servers were offered for about 20 million yuan.2 These were reported unofficial-market offers, not official Nvidia prices or evidence that every transaction occurred at those levels.2

• Domestic substitution remained a compromise, particularly for training. Although the H200 was no longer Nvidia’s newest processor, it retained an advantage over Chinese alternatives for AI training.2 Reports of a possible Chinese approval ceiling below 200,000 H200s, alongside a preference for domestic processors for inference—running already-trained models—suggested selective access rather than unrestricted imports; the ceiling was not a confirmed official policy.2

• Export-compliant alternatives faced a separate supply bottleneck. Bloomberg reported orders for roughly 6 million less-powerful RTX Pro 5500 chips, but buyers expected only about 1% within four months because Nvidia’s production capacity was tight.3 Thus, switching to a US-export-compliant product did not promise prompt access to sufficient computing power.3

Taken together, the reports show that Chinese AI firms’ purchasing demand could not readily translate into usable compute: export rules, Chinese import approvals and manufacturing constraints jointly limited availability, raised costs and forced compromises over processor performance and sourcing.23

As of November 6, Reuters’ November 5 report revealed a previously unreported stockpile of about 600,000 Nvidia H200 chips in Hong Kong awaiting shipment to mainland China—not because buyers lacked demand, but because Beijing was withholding large-scale import approval.2 The report illustrates a two-sided squeeze: US export controls restrict which advanced chips Chinese firms can buy, while China’s own import restrictions limit access even to chips Washington has authorized.2

• US permission did not guarantee Chinese entry. Washington had authorized H200 sales to approved Chinese buyers following US–China trade talks in 2025, but Beijing remained reluctant to deepen dependence on US technology despite rising domestic AI demand.2

• Deliveries were limited, not entirely blocked. Sources said several major Chinese internet companies had each received about 10,000 H200s, with smaller quantities arriving through intermediaries and indirect routes—far short of the Hong Kong stockpile.12 Chinese regulators and Nvidia, Alibaba, ByteDance, DeepSeek and Tencent did not respond to Reuters’ requests for comment, leaving the account without their public confirmation.2

• Unofficial-market prices showed the scarcity premium. An eight-card H200 server was offered for more than 6 million yuan, approximately $900,000 and roughly triple its price a year earlier; export-controlled Blackwell servers were offered for about 20 million yuan.2 These were reported unofficial-market offers, not official Nvidia prices or evidence that every transaction occurred at those levels.2

• Domestic substitution remained a compromise, particularly for training. Although the H200 was no longer Nvidia’s newest processor, it retained an advantage over Chinese alternatives for AI training.2 Reports of a possible Chinese approval ceiling below 200,000 H200s, alongside a preference for domestic processors for inference—running already-trained models—suggested selective access rather than unrestricted imports; the ceiling was not a confirmed official policy.2

• Export-compliant alternatives faced a separate supply bottleneck. Bloomberg reported orders for roughly 6 million less-powerful RTX Pro 5500 chips, but buyers expected only about 1% within four months because Nvidia’s production capacity was tight.3 Thus, switching to a US-export-compliant product did not promise prompt access to sufficient computing power.3

Taken together, the reports show that Chinese AI firms’ purchasing demand could not readily translate into usable compute: export rules, Chinese import approvals and manufacturing constraints jointly limited availability, raised costs and forced compromises over processor performance and sourcing.23